NewsAnalysis
Phishing emails now carry a second message for the AI assistant: hidden text that makes a payment change the assistant's own recommendation
Cybernews reported on October 19, drawing on Malwarebytes' October 12 post and Barracuda's October 7 research, that phishing emails are being written for two readers at once. The human gets a password-protected attachment with the password in the message; the assistant that summarizes the inbox gets instructions hidden in HTML comments, invisible styling, encoded content or zero-width characters, which stay readable to a model that extracts the email's text. In Malwarebytes' example a hidden block told the summarizer to add a priority action, changing a vendor's payment details, so the fraud arrives as the assistant's own advice rather than as a request in a suspicious email. Whether a given tool sees the hidden text depends on how it prepares email content. Both vendors advise checking the original message before acting on any recommendation about money, verifying payment changes through a separate channel, limiting what the assistant may do, and requiring a person to approve payments and vendor changes after reading the underlying request.
Why it matters
The summary is not verification. An assistant that reads untrusted email will repeat what the email tells it, so the control is on the action, a payment or a vendor change, not on the text.